Why Hyper-V for VDI? VIR324

Michael KleefSenior Technical Product Manager

Microsoft

Session Objectives and Takeaways

Windows Server 2008 R2 SP1Dynamic Memory and RemoteFXIs Dynamic Memory just for VDI?

Why is VDI interesting to test?What density is Hyper-V now getting on core workloads?Was that 40% higher density announcement for real?Sure it was. How did we achieve it?

Windows Server 2008 R2 SP1

How does SP1 change the game?

Dynamic MemoryDynamically add/remove memory based on VM demandDelivers 40% increases in VDI density!Supports:

Windows Vista and Windows 7Windows Server 2003, 2008 and 2008 R2

RemoteFXServer side graphics accelerationMore efficient encode/decode USB redirectionSupports:

Windows 7 SP1 Enterprise and Ultimate on Hyper-V 2008 R2 SP1Note: Not for RD Session Host for GPU rendering or USB redirection

*Based on a 3 year amortization and publicly available information as of February 2010.

What does a 40% increase in density mean?

$42/device annual

$21/device annual

$133/ ccu

MS

$63/device

Licensing Comparison of VDI*

Citrix

Microsoft, Citrix and Dell VDI delivers savings – with lowest CAPEX numbers.

Hardware and Hypervisor scale of VDI

Hyper-V 2008 R2 RTM

1000 users

Hyper-V 2008 R2 SP1

1440 users

$320/user2

$222/user2

VMware

1280 users

$242/user1

1Based on: http://www.vmware.com/files/pdf/EMA-VMware-View-StatelessVirtualDesktop-WP.pdf?src=ie9tr2Based on Dell MSRP pricing online as at 3/17/2011 of hardware used in MMS keynote

MS VDI Standard SuiteCitrix XenDesktop

“VDI Edition”

VMware View Premier

Is Dynamic Memory just for VDI?

Short answer: No!• Strong customer success with core infrastructure, web and SQL

workloads“…with Hyper-V and Dynamic Memory, we've been able to consolidate our data center footprint by more than 50 percent, increase server capacity by about 35 percent, and reduce power costs by 33 percent…”

Alan Bourassa, CIO, EmpireCLS Worldwide Chauffeured Services

“…We found that we could increase the number of virtual machines per Hyper-V server by 25 to 50 percent with Dynamic Memory…”

David Feng, IT Director, Sporton International

What workloads work with Dynamic Memory?

Core InfrastructureSQLWeb

Not:Microsoft ExchangeSpecific Applications that check memory @ start and allocateAnything that doesn’t work well with hot add and hot remove

Dynamic Memory for Servers

Demo

Virtual machine-based Remote Desktops (VDI)

Session-based Remote Desktops (RD Session Host

fka Terminal Services)

What is VDI anyway?

Think of VDI like RD Session Host with higher isolation but with less scale and more cost

Why is VDI interesting to test?

What does a typical virtualized environment look like?

Server VirtualizationRelatively easy to see ROI with strong TCO gains81% of customers virtualize anywhere between 1 and 14 guests per server

Source: IDC

Client VirtualizationVery difficult to project ROI with unproven TCO gainsBetween 30 to 120+ guests per serverDisk IO hit especially hard

Core attributes of a good VDI platform

ReliabilityRobustnessCost effectivenessHigh PerformanceDensityRedundancySecurityEase of configurationEase of management

Performance and Density

VDI Capacity Planning and TestingPerformance is very subjective with many variables

CaveatsData provided is based on benchmark results and is not reflective of many real-life deployment considerations:

Is based on ProjectVRC LoginVSI 2.1.2 Medium WorkloadPartially accounts for necessary “cushion” to deal with temporary peaks in resource usage

Piloting critical for performance planningMultiple factors determine actual performance

Variations in hardwareDriver versionsDesktop WorkloadsApplication qualityRoaming Profiles

What we used:HP DL 380 G6 – Dual Quad HT Intel Nehalem, 110GB RAM, 42 disk iSCSI targetDell M610 Blade, Dual Hex core HT Intel Westmere, 96 GB RAM, 16 disk Equallogic SAN’s

Objectives to be determined:How does Dynamic Memory influence VDI scale – what increases are found? What does this do to key indicators?Comparison against Session Virtualization scale on same hardware

What IO bottlenecks do you hit first?

In order, generally that is:Disk IOMemory pressureProcessor

Disk IO is a performance and density related impactMemory is a density impactProcessor is a performance and density related impact

ProcessorRule of thumb: If it doesn’t have SLAT don’t buy it

# of VMs per core/LP is highly dependent on user scenariosApplication specific usage play a big role

Hyper-V supports:1000 VMs per cluster in Clustered scenarios (max of 384 VMs per server)384 VMs per Server in non-Clustered scenarios New! 12 VM’s per Core/Logical Proc

12 VM’s/core is not an architectural limitation but what we have tested and support

SLAT enabled processors provide up to 25% improvement in density

What is Second Level Address Translation (SLAT)? Intel calls it Extended Page Tables (EPT)AMD calls it Nested Page Tables (NPT) or Rapid Virtualization Indexing (RVI)Processor provides two levels of translation

Walks the guest OS page tables directlyNo need to maintain Shadow Page TableNo hypervisor code for demand-fill or flush operations

Resource savingsHypervisor CPU time drops to 2%Roughly 1MB of memory saved per VM

Disk IORule of thumb: SANs are your new best friends

Disk performance is the most critical factor in achieving densityInternal testing showed Windows 7 having lower Disk IO than Windows XP, after boot up

So did ProjectVRC’s recent testing

SAN is of critical importance. Highly recommendedPlenty of cacheConsider de-duplication support especially if persistentDe-duplication allows the benefits of individual images at the cost of differencing diskManaging images on a SAN is way faster and easier than over network (provisioning is faster)We mean real SAN (iSCSI or FC) not NAS across the network…Remember RDS does not require this huge SAN investment…

If you have low complexity requirements:Think about cheaper DAS RAID 0+1 offers better read and write performance than RAID 5Make sure to consider RDS

How does SP1 change the gameStatic Memory

Rule of thumb: More is better

Biggest constraint of upper limit VM density (not performance related)Constrained by:

Available memory slots in serversLargest Available DIMMs

Creates an artificial scale ceiling

Buy as much RAM as you expect to scale the number of VM’sPlan for and allocate at least 1GB per Windows 7 VM on Hyper-V RTM

Memory allocation should be determined by upper maximum limit of running appsAllocate enough RAM to prevent the VM paging to disk1GB actually covers a fair amount of app use….

Also refer to: http://www.microsoft.com/whdc/system/sysperf/Perf_tun_srv-R2.mspx

But Dynamic Memory changes all of the above!!

How does SP1 change the gameDynamic Memory

Rule of thumb: Less is more!!!

Still a constraint of upper limit VM density (not performance related)Buy as much RAM as you expect to scale the number of VM’s

Optimal price/performance curve at 96GB RAM

Plan for and allocate at least 1GB 512MB startup per Windows 7 VM on Hyper-V R2 SP1

Memory allocation will be determined by Dynamic Memory based on running appsActual memory pressure testing in pilot is CRITICALEnsure enough spare capacity to prevent the VM paging to disk

YOU DON’T NEED TO WEAKEN WINDOWS 7 SECURITY TO GET IT TO WORK!

What difference did this make in testing? A lot!

Test Results: LoginVSI and response timeHP DL 380 G6 + iSCSI Storage (42 disks)

Windows 7 64 BitVSI Max was never reached120 VMs achieved with 87GB RAM in useRoaming Profiles not implementedNotice odd spike at end

Test Results: Performance MonitorHP DL 380 G6 + iSCSI Storage (42 disks)

Current Disk Queue Length

(Aggregate of 42 disk Storage target + local

disk usage)

Memory Add Operations % of working set

Peak test load reached

PP Processor %

Test Results: LoginVSI and response timeDell M610 Blade + EqualLogic SAS SAN (16 disks)

Windows 7 32 BitVSI Max was never reached120 VMs achieved with 87GB RAM in useRoaming Profiles not implementedNewer results tool used

Test Results: LoginVSI and response timeDell M610 Blade + EqualLogic SAS SAN (16 disks)

Windows 7 32 BitMemory up slightly on VMProcessor never capped outPlenty of RAM available

% Total Run Time (Total Processor load)

Memory Operations % remaining of working

set

Memory Operations A single VM memory load (no higher than

540MB)

Test Results: LoginVSI and response timeDell M610 Blade + EqualLogic SAS SAN (16 disks)

VSI Max also never reached Easily scaled to 120 VMs with 87GB in useAgain, roaming profiles not implementedNote: Same odd spike at end – uncorrelated to the test data

Test Results: Performance MonitorDell M610 Blade + EqualLogic SAS SAN (16 disks)

Memory Operations % remaining of working set

Peak test load reached

PP Processor %

Current Disk Queue Length

(Aggregate of 16 disk SAN + local disk

usage)

Dell Reference ArchitectureAdding Dynamic Memory

Test Results: LoginVSI8 Dell M610 Blades with XenDesktop

960 VMs deployed using Citrix Provisioning ServicesPVS delivering from Equallogic SSD SANvDisk cache per VM, located on Equallogic SAS SAN

Test Results: LoginVSIOne of eight: Dell M610 Blades with XenDesktop

Memory Operations % remaining of working set

Peak test load reached

PP Processor %

Current Disk Queue Length Scale 1/100

Length never hits beyond 1

How not to do it…The “sum of the parts” considerations

Everything could be so right…Powerful Dell bladesDeployed using Citrix Provisioning ServicesPVS delivering from EqualLogic SSD SANvDisk cache per VM, located on Equallogic SAS SAN

So what caused this mess? Roaming Profiles across “slow” file server and network connection

VDI is complicated and requires careful planning and architecture

What else affects density?

Poor storage architecture – primary candidateAnti VirusRoaming Profiles

Look at AppSense or Citrix to help with this

Slow networking to core infrastructure servicesLack of NIC based TCP offloadingPoorly performing drivers

Conclusion

Hyper-V definitely delivers high scale at low costTotal stack cost of Citrix and Microsoft lower than VMwareDo a POC

Hyper-V is a part of Windows Server 2008 R2You’ll be pleasantly surprised how you don’t have to add custom settings on Hyper-V to get great performanceYou’ll also love how efficient the disk IO is!

Next steps to check out:http://www.citrix.com/xendesktop http://www.microsoft.com/hyperv

Resources

www.microsoft.com/teched

Sessions On-Demand & Community Microsoft Certification & Training Resources

Resources for IT Professionals Resources for Developers

www.microsoft.com/learning

http://microsoft.com/technet http://microsoft.com/msdn

Learning

http://northamerica.msteched.com

Connect. Share. Discuss.

Complete an evaluation on CommNet and enter to win!

© 2011 Microsoft Corporation. All rights reserved. Microsoft, Windows, Windows Vista and other product names are or may be registered trademarks and/or trademarks in the U.S. and/or other countries.The information herein is for informational purposes only and represents the current view of Microsoft Corporation as of the date of this presentation. Because Microsoft must respond to changing market conditions, it should not be interpreted to

be a commitment on the part of Microsoft, and Microsoft cannot guarantee the accuracy of any information provided after the date of this presentation. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS

PRESENTATION.